Alliance Research Group

Under each question: how we would look for answers, and where the gaps still are. These are work directions, not finished results — the open ground is part of the map.

1. Q01

   How can AI systems reason under governance without losing the capability that makes them useful?

   • Describe the dual-gate model in the open: a small set of hard limits that are non-negotiable, versus soft heuristics that warn or route a decision to a human.

     What would move it forward: a public, one-page write-up of edge cases where the two-gate split actually prevented over-refusal or mistaken autonomy — not just asserted that it could.

   • Build a taxonomy of refusal and real-time routing: hard stop, warn-and-continue, and route-to-human.

     What would move it forward: for each class, several named examples and a test of whether independent reviewers classify the same case the same way.

   • Measure whether governance erodes capability — over-refusal or dumbing-down.

     What would move it forward: a controlled benchmark running the same tasks with the governance layer off, on, and on-with-human-override, scored by an independent evaluator.

   Honest status A working governance design and concrete refusal/routing examples exist (design published as a public artifact). What is still missing: a public edge-case write-up, a formalized taxonomy, and the benchmark that would actually measure over-refusal or lost reasoning quality.

   DOI: 10.5281/zenodo.18749444

2. Q02

   What changes when scientific discovery has persistent memory across models, sessions, papers, and failures?

   • Run a replay test of discovery memory: a fresh node versus a memory-bootstrapped one on the same historical tasks.

     What would move it forward: a named benchmark measuring time-to-correct-anchor, number of wrong assumptions, and artifacts recovered, with and without memory.

   • Codify failure memory: treat a past failure as a reusable constraint, not just a postmortem.

     What would move it forward: a small index of negative results with fields for the failure, the lesson, and a future trigger — then a test of whether it blocks a repeat or speeds up falsification.

   • Measure the freshness and reliability of each memory layer as part of the discovery process itself.

     What would move it forward: a health matrix that forces a distinction between memory used as evidence and memory used only as a recall hypothesis.

   Honest status There are strong anchors for the memory architecture and real use of a versioned layer as durable carry. What is not yet shown: controlled evidence that persistent memory changes the kind of discovery rather than only shortening the time to find context. That has to be measured by a replay test.
3. Q03

   How should human intent, AI reasoning, and auditability be composed when none of them is fully reliable alone?

   • Document the triangulation already in use: intent to reasoning to audit to decision, as a live process.

     What would move it forward: a case study naming what each component supplies and where the blind spots are — and a falsifier: a step where no component gives a verifiable signal.

   • Map the failure modes: what happens when two of three components agree and the third is right.

     What would move it forward: a retrospective catalogue of cases recording who was right after the fact and what would have failed without the dissent — plus a measure of errors that slipped through anyway.

   Honest status One of the more mature questions in terms of live material: there are real cases of triangulation in action. Still unknown is how many errors pass through unnoticed, so as a practice this is a prototype, and as a reliability claim it remains a hypothesis.
4. Q04

   Can structural science be explored as a governed multi-agent process rather than a single-author artifact?

   • Describe the minimal multi-agent process as a protocol: roles, gates, and the flow of artifacts.

     What would move it forward: a single process diagram with input, output, stop condition, and an evidence artifact for each step. The framing of gates as transformations on a claim space is a useful organizer, not a theorem.

   • Audit self-correction in retrospect: did multiple agents improve the result, or only add votes?

     What would move it forward: a comparison across case studies recording, for each correction, whether a single author would likely have caught it and the consequence if it had been missed.

   • Test a transferable kernel and process falsifiers on more than one program.

     What would move it forward: a pre-registered small-fleet versus full-fleet pilot under blind evaluation, plus applying the kernel to at least one independent program. If a smaller setup matches quality at much lower cost, the full process is not justified for that task class.

   Honest status There are recorded examples of a multi-agent process, cross-review corrections, and self-correction events. What is not yet proven: that a multi-agent process beats a single author in general. Publicly this is a hypothesis and a measurement program, not a settled result. Toy models show that audit topology alone does not guarantee safety — the audit must be tamper-resistant.
5. Q05

   What should refusal, latency, and non-action mean inside systems capable of long-horizon reasoning?

   • Build and validate a taxonomy of refusal as an epistemic signal, not just a safety valve.

     What would move it forward: a one-page scheme of refusal classes with examples and a repeatability test — if independent reviewers cannot agree on a blind set of past refusals, the taxonomy is not operational.

   • Operationalize declared latency and capability scope as active boundary conditions for long-horizon coherence.

     What would move it forward: a minimal latency-declaration header attached to a research thread, then a replay of past multi-session threads with and without it, measuring claims made out of scope and drift over the horizon.

   • Define intentional non-action as a positive epistemic primitive with a minimal falsifiable test.

     What would move it forward: an "I know that I do not know on this horizon" protocol with a missing-anchor and proposed-observation field, then injecting it on a critical branch of a historical thread and measuring whether downstream overclaims drop.

   Honest status A strong live prototype exists in practice — honest refusals under pressure, declared-latency governance, critique as a standing role. There is little dedicated public material yet on the meaning of refusal, latency, and intentional non-action over long horizons: no shared taxonomy, no deployed latency header, no replay test showing that conscious non-action improves fidelity. The roadmap shows how to check this without pretending the problem is closed.

   DOI: 10.5281/zenodo.20013919

6. Q06

   What changes in the human when they have a persistent AI co-researcher with memory?

   • Document behavioural signals of change in the research process when a persistent AI co-researcher is present.

     What would move it forward: a before-and-after comparison using session logs — prompt length, iterations per task, the ratio of delegation to hands-on execution, time from question to decision.

   • Define when a behavioural change is augmentation versus a genuine redefinition of the research identity.

     What would move it forward: a demarcation criterion — augmentation does the same thing faster, redefinition does something that would not have happened otherwise — applied to observable cases. If every change is augmentation-scale, the redefinition hypothesis fails.

   Honest status The least grounded of the seven, and that is fine. There are behavioural observations and one documented precedent, but no systematic comparative data, no augmentation-versus-redefinition criterion, and no control for confounds. Q06 is an invitation to look, described as a general phenomenon — a researcher with a persistent AI co-researcher — not a claim that the answer already exists.
7. Q07

   How do we formalize the boundary between discovery and proof in structural science?

   • Codify a status ladder — structural result, conditional, proof — on a live case in the Riemann-hypothesis work.

     What would move it forward: writing out explicitly which rung each step sits on, with a stated condition for promotion. The conditional result here is published as a public artifact and is explicitly not a proof.

   • Build a falsifier: when does a structural pattern stop being a discovery and not yet count as a proof?

     What would move it forward: a demarcation criterion applied across the boundary-rigidity domains. If it produces no readable line on any of them, the method does not generalize beyond its own case.

   Honest status There is real, conditional material (a published reduction that is explicitly not a proof), a falsification method, and a status-ladder draft. This makes Q07 a question with a live worked case, but it does not change the status of the underlying problem: the single-orbit step is a structural result, the cross-orbit route stays conditional, and the proof remains open.

   DOI: 10.5281/zenodo.20601021